Haoran Bai

I am also very interested in the offset-fidelity loss presented in "Understanding Deformable Alignment in Video Super-Resolution". It is awesome to use optical flow to stabilize the training process. Could...

A feasible solution is to use 3D software (e.g., R3DS Wrap) to find the face indices of the eyebrows, and then find the corresponding UV coordinates based on these face...

Maybe you can try `git clone -b v0.6.2 https://github.com/facebookresearch/pytorch3d`. This version supports Python 3.7 and PyTorch 1.7

We have provided the 68 3D landmarks on the HiFi3D++ mesh [[here]](https://github.com/csbhr/FFHQ-UV/blob/main/README_ckp_topo.md#the-introduction-of-these-topology-assets). One can find the corresponding pixel position in the UV maps based on the UV coordinates. There is...

Are you referring to using Deep3D to regress the latent code `z` of the texture decoder? This is possible. Using Deep3D, a self-supervised training strategy, the more high-quality training data,...

Please refer to [here](https://github.com/csbhr/FFHQ-UV/blob/main/README_rgb_fitting.md#reproducing-realy-benchmark-results) for details of reproducing REALY benchmark results.

To calculate the evaluation metrics of REALY benchmark, please follow the instructions of the [official repository](https://github.com/czh-98/REALY). Downloading benchmark data requires signing an **Agreement**.

There are two possible ways to achieve this: 1. Use the interactive nonrigid registration tool to align the HiFi3D++ UVmap to the FLAME topology; 2. Recreate the FLAME topology assets,...